Conventional memory devices have a standard interface consisting of separate address, data and control pins. For example, one version of a synchronous dynamic random access memory (SDRAM has twelve address pins, two multiplexed address and control pins, seven control pins and sixteen data pins. This approach offers a great deal of flexibility since computer systems vary greatly in their memory requirements. In particular, the bandwidth of memory systems using SDRAMs can easily be increased by adding another SDRAM in parallel to the existing SDRAMs, thereby increasing the width of the memory bus.
The tradeoff for this flexibility is, however, an increase in layout space which leads to an increase in manufacturing cost. Separate traces need to be run for each pin of each SDRAM. Therefore, it is sometimes cost prohibitive to use SDRAMS for wide memory systems.
One approach to lower the cost of expanding memory is to use memory devices which multiplex address, control and data information on the same pins. For example, some memory devices have a set of generic interface pins which connect to a high-speed, synchronous bus. Communication over the bus is accomplished by a series of packets which conform to a predefined packet protocol. Usually the packet protocol is fairly sophisticated and has a complete command set. For example, DRAMS conforming to the RAMBUS.TM. interface communicate using a protocol in which each packet consists of six bytes transmitted sequentially over a high-speed bus known as a "Channel." In this manner, the packets encapsulate all address, control and data information.
Because of the efficient use of generic interface pins, a packet protocol reduces the required number of pins to approximately 30. However, this has the disadvantage of decreasing effective data bandwidth, because only a portion of the total bus bandwidth is available for data (the rest of the bandwidth is reserved for address and control information).
Another method for reducing the cost associated with increasing total memory bandwidth, without decreasing effective data bandwidth, is to provide a second high-speed bus specifically for communicating data In this approach, address and control information is communicated over a unidirectional high-speed address/control bus while data is communicated over a bidirectional high-speed data bus. Both communications conform to a predefined packet protocol. This approach has the benefits of reducing the total pin count (although not as much as the RAMBUS.TM. protocol described above), yet has the added benefit that only the data bus needs to be duplicated when the width of the memory system is increased.
Both approaches described above offer advantages over traditional memory architectures in terms of increased data retrieval bandwidth It is difficult, however, to implement systems having both fine granularity and large memory depth using such devices. What is needed is a memory architecture which supports increased bandwidth, fine granularity, and large memory arrays.